Can a Photograph Fool Your Face? Deepfakes, Spoofing and Biometric Security

Blurs.co.uk
10 Min Read
Photo: PO Phot Derek Wade / Ministry of Defence, Wikimedia Commons. Open Government Licence v1.0.
Summarize with AI
A person holding a mobile phone, illustrating smartphone facial authentication and biometric security.
Photo: PO Phot Derek Wade / Ministry of Defence, Wikimedia Commons. Open Government Licence v1.0.

Facial recognition is often presented as a convenient way to prove who you are. A glance at a phone can unlock a screen, a selfie can help open a bank account, and a short video can be used to confirm an identity online. Yet these systems face a deceptively simple problem: how can a computer know that the face in front of its camera belongs to a real, present person rather than a photograph, a replayed video or a digitally generated imitation?

The challenge is known in biometric security as a presentation attack. It is not the same as finding a lookalike or tricking a search algorithm into matching the wrong person. Instead, an attacker tries to fool the camera or its software into accepting an artificial representation of the genuine user’s face.

When a face is only an image

Basic facial unlocking systems may rely on a conventional two-dimensional camera. They capture an image, compare it with the face enrolled when the device was set up and unlock if the similarity score passes a threshold. Such a system can be convenient, but a flat image contains limited information about depth and physical presence.

A printed photograph or a picture displayed on another screen may therefore be enough to fool some implementations. A replayed video can add movement, while more sophisticated attacks may use masks or manipulated camera feeds. The effectiveness of an attack depends on the device, its sensors, the software, the security settings and the attacker’s access.

Not all facial recognition is equally vulnerable. Some systems use infrared sensors, depth mapping or structured light to measure three-dimensional features. Others combine facial matching with device security hardware and additional authentication checks. These approaches can make spoofing harder, although no single feature should be treated as an absolute guarantee.

What recent smartphone testing found

In April 2026, the UK consumer organisation Which? reported the results of laboratory testing of 208 smartphones assessed since October 2022. It said 133 devices, or 64 per cent of those tested, had facial unlocking that could be fooled by a printed two-dimensional photograph. The findings included models from a range of well-known manufacturers.

This is a warning about the devices in that test sample, not proof that 64 per cent of every smartphone in use is vulnerable. The results also do not mean that every phone using facial recognition is unsafe for every purpose. Some systems are explicitly designed as a convenience feature rather than a high-assurance security method, and manufacturers may prevent lower-security face unlock from authorising banking or payment functions.

Which? noted differences between basic 2D systems and more advanced approaches using depth information. Its findings underline why consumers should check the security class and limitations of a particular device rather than assume that all features labelled face recognition offer the same protection.

Deepfakes make the problem more complicated

Generative artificial intelligence has made it easier to create convincing synthetic images, voices and video. In a remote identity check, an attacker may attempt to present a manipulated video stream or an animated digital face instead of appearing directly in front of the camera.

A convincing deepfake does not automatically defeat a well-designed biometric system. The attack must also bypass the application’s capture process, the device’s security controls and any anti-spoofing checks. But the growing quality of synthetic media raises the stakes for services that rely on a selfie or short video as their main evidence that an applicant is genuine.

It is useful to distinguish between two separate questions. Face matching asks whether the person or image resembles the enrolled identity. Liveness or presentation-attack detection asks whether the camera is seeing a genuine person in the expected way. A system can perform one task well and still fail at the other.

How liveness detection works

Some systems ask a user to move their head, blink, look in a particular direction or respond to a random prompt. These challenge-response methods can make simple replay attacks more difficult because a pre-recorded clip may not respond correctly to an unpredictable instruction.

Passive systems attempt to detect signs of a real face without asking the user to perform a deliberate action. They may analyse subtle image characteristics, texture, reflections, depth cues or movement. More advanced approaches can combine several sensors or signals, making it harder for a single photograph or video to satisfy all checks.

However, liveness detection is not a magic shield. A poorly designed prompt can be anticipated, a passive detector may be trained on limited examples, and a system that works in a laboratory may struggle under different lighting, camera quality or network conditions. Security tests need to include attacks that were not part of the original training data.

What standards and government guidance say

The US National Institute of Standards and Technology published a detailed evaluation of software-based presentation-attack detection in 2023. It assessed 82 prototype algorithms from 45 developers, examining the trade-off between detecting attacks and incorrectly rejecting genuine users. NIST stressed that its evaluation covered specified image-based methods and did not test every possible hardware or interactive approach.

In Britain, government guidance on identity proofing and verification says providers using biometric checks should consider false matches and false non-matches, use algorithms tested against recognised benchmarks, and ensure systems can detect spoofing. It also refers to enhanced liveness checks and independent testing by appropriately accredited laboratories.

These principles matter because identity verification is not a single algorithm. It is a chain involving image capture, document checks, data transmission, biometric comparison, anti-spoofing measures and the organisation’s decision about what to do with the result. A weakness in any link can undermine the overall process.

Why biometric data creates a lasting risk

A compromised password can be changed. A compromised face cannot. Facial templates are not necessarily stored as ordinary photographs, and well-designed systems may keep sensitive matching data in secure hardware or on the user’s device. Nevertheless, organisations need to explain what is collected, where it is stored, who can access it and how long it is retained.

There is also a risk of function creep. A face scan collected to unlock a phone should not quietly become a tool for tracking a person’s activity across unrelated services. A business collecting a selfie to verify an account should not assume that it can reuse the image for advertising, profiling or unrelated identification.

Practical steps for users

Consumers should check whether a phone’s face-unlock feature is described as suitable for sensitive applications or only for convenience. If the manufacturer warns that a photograph or similar-looking person may unlock the device, a strong PIN, password or fingerprint method may be a safer choice. Sensitive apps can often be protected with an additional authentication step.

For online identity checks, users should access the service through its official app or website, avoid sending identity videos through unsolicited links and be alert to unexpected requests for selfies or documents. Organisations should provide alternative ways to verify identity when a user cannot complete a face scan.

The real test is trust, not novelty

Facial recognition can reduce friction and help prevent some forms of impersonation. But a system should not be trusted simply because it looks futuristic or uses artificial intelligence. Its performance must be independently tested, its limitations explained and its fallback procedures made clear.

The most secure systems recognise that a face image is evidence, not proof on its own. A reliable identity check combines several safeguards and remains prepared for the possibility that the camera, the software or the underlying assumptions may be wrong.

Sources and further reading

Share This Article
Leave a Comment

Leave a Reply